STOCHASTIC ANALYSIS OF GAUSSIAN PROCESSES VIA
FREDHOLM REPRESENTATION
arXiv:1410.2230v1 [math.PR] 8 Oct 2014
TOMMI SOTTINEN AND LAURI VIITASAARI
Abstract. We show that every separable Gaussian process with integrable variance function admits a Fredholm representation with respect
to a Brownian motion. We extend the Fredholm representation to a
transfer principle and develop stochastic analysis by using it. In particular, we prove an Itˆ
o formula that is, as far as we know, the most
general Malliavin-type Itˆ
o formula for Gaussian processes so far. Finally, we give applications to equivalence in law and series expansions
of Gaussian processes.
1. Introduction
The stochastic analysis of Gaussian processes that are not semimartingales is challenging. One way to overcome the challenge is to represent the
Gaussian process under consideration, X say, in terms of a Brownian motion
and then develop a transfer principle so that that the stochastic analysis can
be done in the “Brownian level” and then transfered back into the level of
X.
One of the most studied representation in terms of a Brownian motion
is the so-called Volterra representation. A Gaussian Volterra process is a
process that can be represented as
Z t
K(t, s) dWs ,
t ∈ [0, T ],
(1.1)
Xt =
0
where W is a Brownian motion and K ∈ L2 ([0, T ]2 ). Here the integration
goes only upto t, hence the name “Volterra”. This Volterra nature is very
convenient: it means that the filtration of X is included in the filtration of
the underlying Brownian motion W . Gaussian Volterra processes and their
stochastic analysis has been studied, e.g., in [2] and [7], just to mention
few. Apparently, the most famous Gaussian process admitting Volterra
representation is the fractional Brownian motion and its stochastic analysis
indeed has been developed mostly by using its Volterra representation, see
e.g. [12] and references therein.
Date: October 9, 2014.
2010 Mathematics Subject Classification. Primary 60G15; Secondary 60H05, 60H07,
60H30.
Key words and phrases. Equivalence in law; Gaussian processes; Itˆ
o formula; Malliavin
calculus; representation of Gaussian processes; series expansions; stochastic analysis.
Lauri Viitasaari was partially funded by Emil Aaltonen Foundation.
1
2
SOTTINEN AND VIITASAARI
In discrete finite time the Volterra representation (1.1) is nothing but the
Cholesky lower-triangular factorization of the covariance of X, and hence
every Gaussian process is a Volterra process. In continuous time this is not
true, see Example 3.1 in Section 3.
There is a more general representation than (1.1) by Hida, see [9, Theorem
4.1]. However, this Hida representation includes possibly infinite number of
Brownian motions. Consequently, it seems very difficult to apply the Hida
representation to build a transfer principle needed by stochastic analysis.
Moreover, the Hida representation is not quite general, e.g., Example 3.1 in
Section 3 does not admit a Hida representation.
The problem with the Volterra representation (1.1) is the Volterra nature
of the kernel K, as far as generality is concerned. Indeed, if one considers
Fredholm kernels, i.e., kernels where the integration is over the entire interval
[0, T ] under consideration, one obtains generality. A Gaussian Fredholm
process is a process that admits the Fredholm representation
Z T
(1.2)
Xt =
KT (t, s) dWs ,
t ∈ [0, T ],
0
where W is a Brownian motion and KT ∈ L2 ([0, T ]2 ). In this paper we show
that every separable Gaussian process with integrable variance function admits the representation (1.2). The price we have to pay for this generality
is twofold:
(i) The process X is generated, in principle, from the entire path of
the underlying Brownian motion W . Consequently, X and W do
not necessarily generate the same filtration. This is unfortunate in
many applications.
(ii) In general the kernel KT depends on T even if the covariance R
didn’t, and consequently the derived operators also depend on T .
This is why we use the cumbersome notation of explicitly stating
out the dependence when there is one. In stochastic analysis this
dependence on T seems to be a minor inconvenience, however. Indeed, even in the Volterra case as examined, e.g., by Al`os, Mazet
and Nualart [2], one cannot avoid the dependence on T in the transfer principle. Of course, for statistics, where one would like to let
T tend to infinity, this is a major inconvenience.
Let us note that the Fredholm representation has already been used, without proof, in [3], where the H¨older continuity of Gaussian processes was
studied.
The paper is organized as follows:
Section 2 contains some preliminaries on Gaussian processes and isonormal Gaussian processes and related Hilbert spaces.
Section 3 provides the proof of the main theorem of the paper: the Fredholm representation.
FREDHOLM REPRESENTATION
3
In Section 4 we extend the Fredholm representation to a transfer principle
in three contexts of growing generality: First we prove the transfer principle
for Wiener integrals in Subsection 4.1, then we use the transfer principle to
define the multiple Wiener integral in Subsection 4.2, and finally, in Subsection 4.3 we prove the transfer principle for Malliavin calculus, thus showing
that the definition of multiple Wiener integral via the transfer principle done
in Subsection 4.2 is consistent with the classical definitions involving Brownian motion or other Gaussian martingales. Indeed, classically one defines
the multiple Wiener integrals by either building an isometry with removed
diagonals or by spanning higher chaoses by using the Hermite polynomials.
In the general Gaussian case one cannot of course remove the diagonals, but
the Hermite polynomial approach is still valid. We show that this approach
is equivalent to the transfer principle. In Subsection 4.3 we also prove an
Itˆo formula for general Gaussian processes. This Itˆo formula is, as far as
we know, the most general version for Gaussian processes existing in the
literature.
Finally, in Section 5 we show the power of the transfer principle in two
applications: in Subsection 5.1 the transfer principle is applied to the question of equivalence of law of general Gaussian processes and in Subsection
5.2 the transfer principle is used to provide series expansions for general
Gaussian processes.
2. Preliminaries
Our general setting is as follows: Let T > 0 be a fixed finite time-horizon
and let X = (Xt )t∈[0,T ] be a Gaussian process with covariance R that may
or may not depend on T . Without loss of any interesting generality we
assume that X is centered. We also make the very weak assumption that
X is separable in the sense of following definition.
Definition 2.1 (Separability). The Gaussian process X is separable if the
Hilbert space L2 (Ω, σ(X), P) is separable.
Example 2.1. If the covariance R is continuous, then X is separable. In
particular, all continuous Gaussian processes are separable.
Definition 2.2 (Associated operator). For a kernel Γ ∈ L2 ([0, T ]2 ) we associate an operator on L2 ([0, T ]), also denoted by Γ, as
Z T
f (s)Γ(t, s) ds.
Γf (t) =
0
Definition 2.3 (Isonormal process). The isonormal process associated with
X, also denoted by X, is the Gaussian family (X(h), h ∈ HT ), where the
Hilbert space HT = HT (R) is generated by the covariance R as follows:
(i) indicators 1t := 1[0,t) , t ≤ T , belong to HT .
(ii) HT is endowed with the inner product h1t , 1s iHT := R(t, s).
4
SOTTINEN AND VIITASAARI
Definition 2.3 states that X(h) is the image of h ∈ HT in the isometry
that extends the relation
X (1t ) := Xt
linearly. Consequently, we can define:
Definition 2.4 (Wiener integral). X(h) is the Wiener integral of the element h ∈ HT with respect to X. We shall also denote
Z T
h(t) dXt := X(h).
0
Remark 2.1. Eventually, all the following will mean the same:
Z T
Z T
h(t) δXt = IT,1 (h) = IT (h).
h(t) dXt =
X(h) =
0
0
Remark 2.2. The Hilbert space HT is separable if and only if X is separable.
Remark 2.3. Due to the completion under the inner product h·, ·iHT it
may happen that the space HT is not a space of functions, but contains
distributions, cf. [16] for the case of fractional Brownian motions with Hurst
index bigger than half.
Definition 2.5. The function space HT0 ⊂ HT is the space of functions
that can be approximated by step-functions on [0, T ] in the inner product
h·, ·iHT .
Example 2.2. If the covariance R is of bounded variation, then HT0 is the
space of functions f satisfying
Z TZ T
|f (t)f (s)| |R|(ds, dt) < ∞.
0
0
Remark 2.4. Note that it may be that f ∈ HT0 but for some T ′ < T we
have f 1T ′ 6∈ HT0′ , cf. [4] for an example with fractional Brownian motion
with Hurst index less than half. For this reason we keep the notation HT
instead of simply writing H . For the same reason we include the dependence
of T whenever there is one.
3. Fredholm Representation
Theorem 3.1 (Fredholm representation). Let X = (Xt )t∈[0,T ] be a separable
centered Gaussian process. Then there exists a kernel KT ∈ L2 ([0, T ]2 ) and
a Brownian motion W = (Wt )t≥0 , independent of T , such that
Z T
KT (t, s) dWs
(3.1)
Xt =
0
if and only if the covariance R of X satisfies the trace condition
Z T
R(t, t) dt < ∞.
(3.2)
0
FREDHOLM REPRESENTATION
5
The representation (3.1) is unique in the sense that any other represen˜ T , say, is connected to (3.1) by a unitary operator U
tation with kernel K
2
˜ T = U KT . Moreover, one may assume that KT is
on L ([0, T ]) such that K
symmetric.
Proof. Let us first remark that (3.2) is precisely what we need to invoke the
Mercer’s theorem and take square root in the resulting expansion.
Now, by the Mercer’s theorem we can expand the covariance function R
on [0, T ]2 as
∞
X
λTi eTi (t)eTi (s),
(3.3)
R(t, s) =
i=1
(λTi )∞
i=1
(eTi )∞
i=1
where
and
covariance operator
are the eigenvalues and the eigenfunctions of the
RT f (t) =
Z
T
f (s)R(t, s) ds.
0
2
Moreover, (eTi )∞
i=1 is an orthonormal system on L ([0, T ]).
Now, RT , being a covariance operator, admits a square root operator KT
defined by the relation
Z T
Z T
T
T
KT eTi (s)KT eTj (s) ds
ei (s)RT ej (s) ds =
(3.4)
eTi
0
0
eTj .
for all
and
Now, condition (3.2) means that RT is trace class and,
consequently, KT is Hilbert–Schmidt. In particular, KT is a compact operator. Therefore, it admits a kernel. Indeed, a kernel KT can be defined by
using the Mercer expansion (3.3) as
∞ q
X
λTi eTi (t)eTi (s).
(3.5)
KT (t, s) =
i=1
This kernel is obviously symmetric. Now, it follows that
Z T
KT (t, u)KT (s, u) du,
R(t, s) =
0
and the representation (3.1) follows from this.
Finally, let us note that the uniqueness upto a unitary transformation is
obvious from the square-root relation (3.4).
Example 3.1. Let us consider the following very degenerate case: Suppose
Xt = f (t)ξ, where f is deterministic and ξ is a standard normal random
variable. Suppose T > 1. Then
Z T
f (t)1[0,1) (s) dWs .
(3.6)
Xt =
0
So, KT (t, s) = f (t)1[0,1) (s). Now, if f ∈ L2 ([0, T ]), then condition (3.2) is
satisfied and KT ∈ L2 ([0, T ]2 ). On the other hand, even if f ∈
/ L2 ([0, T ]) we
can still write X in form (3.6). However, in this case the kernel KT does
not belong to L2 ([0, T ]2 ).
6
SOTTINEN AND VIITASAARI
4. Transfer Principle and Stochastic Analysis
4.1. Wiener Integrals. The following Theorem 4.1 is the transfer principle
in the context of Wiener integrals. The same principle extends to multiple
Wiener integrals and Malliavin calculus later in the following subsections.
Recall that for any kernel Γ ∈ L2 ([0, T ]2 ) its associated operator on
is
Z T
f (s)Γ(t, s) ds.
Γf (t) =
L2 ([0, T ])
0
Definition 4.1 (Adjoint associated operator). The adjoint associated operator Γ∗ of a kernel Γ ∈ L2 ([0, T ]2 ) is defined by linearly extending the
relation
Γ∗ 1t = Γ(t, ·).
Remark 4.1. The name and notation of “adjoint” for KT∗ comes from Al`os,
Mazet and Nualart [2] where they showed that in their Volterra context KT∗
admits a kernel and is an adjoint of KT in the sense that
Z T
Z T
∗
KT f (t) g(t) dt =
g(t) KT f (dt).
0
0
for step-functions f and g belonging to L2 ([0, T ]). It is straightforward to
check that this statement is valid also in our case.
Example 4.1. Suppose the kernel Γ(·, s) is of bounded variation for all s
and that f is nice enough. Then
Z T
∗
f (t)Γ(dt, s).
Γ f (s) =
0
Theorem 4.1 (Transfer principle for Wiener integrals). Let X be a separable
centered Gaussian process with representation (3.1) and let f ∈ HT . Then
Z T
Z T
KT∗ f (t) dWt .
f (t) dXt =
0
0
Proof. Assume first that f is an elementary function of form
f (t) =
n
X
ak 1Ak
k=1
for some disjoint intervals Ak = (tk−1 , tk ]. Then the claim follows by the very
definition of the operator KT∗ and Wiener integral with respect to X together
with representation (3.1). Furthermore, this shows that KT∗ provides an
isometry between HT and L2 ([0, T ]). Hence HT can be viewed as a closure of
elementary functions with respect to kf kHT = kKT∗ f kL2 ([0,T ]) which proves
the claim.
FREDHOLM REPRESENTATION
7
4.2. Multiple Wiener Integrals. The study of multiple Wiener integrals
go back to Itˆo [11] who studied the case of Brownian motion. Later Huang
and Cambanis [10] extended to notion to general Gaussian processes. Dasgupta and Kallianpur [6, 5] and Perez-Abreu and Tudor [15] studied multiple
Wiener integrals in the context of fractional Brownian motion. In [6, 5] a
method that involved a prior control measure was used and in [15] a transfer principle was used. Our approach here extends the transfer principle
method used in [15].
We begin by recalling multiple Wiener integrals with respect to Brownian
motion and then we apply transfer principle to generalize the theory to
arbitrary Gaussian process.
Let f be a elementary function on [0, T ]p that vanishes on the diagonals,
i.e.
n
X
ai1 ...ip 1∆i1 ×···×∆ip ,
f=
i1 ,...,ip =1
where ∆k := [tk−1 , tk ) and ai1 ...ip = 0 whenever ik = iℓ for some k 6= ℓ. For
such f we define the multiple Wiener integral as
Z T
Z T
W
f (t1 , . . . , tp ) δWt1 · · · δWtp
···
IT,p (f ) :=
0
0
:=
n
X
ai1 ...ip ∆Wt1 · · · ∆Wtp ,
i1 ,...,ip =1
where we have denoted ∆Wtk := Wtk − Wtk−1 . For p = 0 we set I0W (f ) = f.
W has the properties
Obviously, the operator IT,p
W (cf ) = cI W (f ) for any c ∈ R,
(i) IT,p
T,p
W (f + g) = I W (f ) + I W (g),
(ii) IT,p
T,p
T,p
W (f ) = I W (f˜), where f˜ is the symmetrization of f, i.e.
(iii) IT,p
T,p
1 X
f˜(t1 , . . . , tp ) :=
f (tσ(1) , . . . , tσ(p) ).
p! σ
Here the summations runs over all permutations σ of {1, . . . , p}.
Since f vanishes at the diagonals we have also the following properties:
h
i
W (f )I W (g) = 0, if p 6= q.
(iv) E IT,p
T,q
h
i
W (f )I W (g) = p!hf˜, g
(v) E IT,p
˜iL2 ([0,T ]p ) ,
T,p
i
h
2
W (f )2 = p!kf˜k2
(vi) E IT,p
L2 ([0,T ]p ) ≤ p!kf kL2 ([0,T ]p ) .
Now, it can be shown that elementary functions that vanish on the diagW to the
onals are dense in L2 ([0, T ]p ). Thus, one can extend the operator IT,p
space L2 ([0, T ]p ) by the isometric property (vi). This extension is called the
multiple Wiener integral with respect to the Brownian motion
8
SOTTINEN AND VIITASAARI
W (f ) can be understood as a multiple
Remark 4.2. It is well known that IT,p
of iterated Ito integral if and only if f (t1 , . . . , tp ) = 0 unless t1 ≤ · · · ≤ tp .
In this case we have
Z t2
Z T Z tp
W
f (t1 , . . . , tp ) dWt1 · · · dWtp .
···
IT,p (f ) = p!
0
0
0
For the case of Gaussian processes that are not martingales this fact is totally
useless.
Let us then consider the multiple Wiener integrals IT,p for a general
Gaussian process X. We define the multiple integrals IT,p and JT,p by using
the transfer principle in Definition 4.3 below and later argue that this is the
“correct” way of defining them. So, let X be a centered Gaussian process
on [0, T ] with covariance R and representation (3.1) with kernel KT .
Definition 4.2 (p-fold adjoint associated operator). Let KT be the kernel
in (3.1) and let KT∗ be its adjoint associated operator. Define
∗
KT,p
:= (KT∗ )⊗p .
(4.1)
In the same way, define
HT,p := HT⊗p
and
0
HT,p
:= (HT⊗p )0 .
Definition 4.3. Let X be a centered Gaussian process with representation
(3.1) and let f ∈ HT,p . Then
W
∗
IT,p (f ) := IT,p
KT,p
f
Note that with this definition properties (i)–(vi) are obviously satisfied.
The following example should further convince the reader that this is indeed
the correct definition.
Example 4.2. Let p = 2 and let h = h1 ⊗ h2 , where both h1 and h2 are
step-functions. Then
∗
(KT,2
h)(x, y) = (KT∗ h1 )(x)(KT∗ h2 )(y)
and
W
∗
IT,2
KT,2
f
Z
Z T
∗
KT h1 (v)dWv ·
=
0
0
T
KT∗ h2 (u)dWu − hKT∗ h1 , KT∗ h2 iL2 ([0,T ])
= X(h1 )X(h2 ) − hh1 , h2 iHT
as supposed to by analog of the Gaussian martingale case.
Recall next the Hermite polynomials:
Hp (x) :=
(−1)p 1 x2 dp − 1 x2 e2
e 2
.
p!
dxp
The following is an extension of well-known product formula for general
Gaussian processes with product integrands. For the case of Gaussian martingales see e.g. Proposition 1.1.4 of [14]. For general Gaussian process,
FREDHOLM REPRESENTATION
9
it seem that this extension is new. Indeed, usually multiple Wiener integrals for more general Gaussian processes are defined as the closed linear
space generated by Hermite polynomials. In our case we define integrals via
transfer principle, and the next proposition shows the connection to Hermite
polynomials.
Proposition 4.1. Let Hp be the pth Hermite polynomial and let h ∈ HT .
Then
X(h)
p
⊗p
= p!khkHT Hp
IT,p h
khkHT
Proof. First note that without loss of generality we can assume khkHT = 1.
Now by the definition of multiple Wiener integral with respect to X we have
W
∗
KT,p
h⊗p ,
IT,p h⊗p = IT,p
where
∗
KT,p
h⊗p = (KT∗ h)⊗p .
Consequently, by [14, Proposition 1.1.4] we obtain
IT,p h⊗p = p!Hp (W (KT∗ h))
which implies the result together with Theorem 4.1.
Proposition 4.1 extends to the following product formula, which is also
well-known in the Gaussian martingale case, but apparently new for general Gaussian processes. Again, the proof is straightforward application of
transfer principle.
Proposition 4.2. Let f ∈ HT,p and g ∈ HT,q . Then
p∧q X
p q
˜ KT ,r g),
r!
(4.2)
IT,p (f )IT,q (g) =
IT,p+q−2r (f ⊗
r
r
r=0
where
∗
∗
˜ KT ,r g = KT,p
˜ r KT,q
f⊗
f⊗
g.
Proof. The proof follows directly from the definition of IT,p (f ) and [14,
Proposition 1.1.3].
Remark 4.3. In the literature multiple Wiener integrals are usually defined
as the closed linear space spanned by Hermite polynomials. In such a case
Proposition 4.1 is clearly true by the very definition. Furthermore, one has
a multiplication formula (see e.g. [13])
p∧q X
p q
˜ HT ,r g),
r!
IT,p (f )IT,q (g) =
IT,p+q−2r (f ⊗
r
r
r=0
˜ HT ,r g denotes symmetrization of tensor product
where f ⊗
f ⊗HT ,r g =
∞
X
hf, ei1 ⊗ . . . ⊗ eir iH ⊗r ⊗ hg, ei1 ⊗ . . . ⊗ eir iH ⊗r .
i1 ,...,ir =1
T
T
10
SOTTINEN AND VIITASAARI
and {ek , k = 1, . . .} is a complete orthonormal basis of the Hilbert space
HT . Clearly, by Proposition 4.1, both formulas coincide.
4.3. Malliavin calculus and Skorohod integrals. We begin by recalling
some basic facts on Malliavin calculus.
Definition 4.4. Denote by S the space of all smooth random variables of
the form
F = f (X(ϕ1 ), · · · , X(ϕn )),
ϕ1 , · · · , ϕn ∈ HT ,
where f ∈ Cb∞ (Rn ) i.e. f and all its derivatives are bounded. The
derivative DT = DTX of F is an element of L2 (Ω; HT ) defined by
n
X
Malliavin
∂i f (X(ϕ1 ), · · · , X(ϕn ))ϕi .
DT F =
i=1
In particular, DT Xt = 1[0,t] .
Definition 4.5. Let D1,2 = D1,2
X be the Hilbert space of all square integrable
Malliavin differentiable random variables defined as the closure of S with
respect to norm
kF k21,2 = E |F |2 + E kDT F k2HT .
The divergence operator δT is defined as the adjoint operator of the Malliavin derivative DT .
Definition 4.6. The domain Dom δT of the operator δT is the set of random
variables u ∈ L2 (Ω; HT ) satisfying
EhDT F, uiH ≤ cu kF kL2
T
for any F ∈ D1,2 and some constant cu depending only on u. For u ∈ Dom δT
the divergence operator δT (u) is a square integrable random variable defined
by the duality relation
E [F δT (u)] = EhDT F, uiHT
for all F ∈ D1,2 .
Remark 4.4. It is well-known that D1,2 ⊂ Dom δT .
We use the notation
δT (u) =
Z
T
us δXs .
0
Theorem 4.2 (Transfer principle for Malliavin calculus). Let X be a separable centered Gaussian process with Fredholm representation (3.1). Let DT
and δT be the Malliavin derivative and the Skorohod integral with respect to
X on [0, T ]. Similarly, let DTW and δTW be the Malliavin derivative and the
Skorohod integral with respect to the Brownian motion W of (3.1) restricted
on [0, T ]. Then
δT = δTW KT∗
and
KT∗ DT = DTW .
FREDHOLM REPRESENTATION
11
Proof. The proof follows directly from transfer principle and the isometry
provided by KT∗ with same arguments as in [2]. Indeed, by isometry we have
HT = (KT∗ )−1 (L2 ([0, T ])),
where (KT∗ )−1 denotes the pre-image, which implies that
2
D1,2 (HT ) = (KT∗ )−1 (D1,2
W (L ([0, T ])))
which justifies KT∗ DT = DTW . Furthermore, we have relation
Ehu, DT F iHT = EhKT∗ u, DTW F iL2 ([0,T ])
for any smooth random variable F and u ∈ L2 (Ω; HT ). Now following [2]
we obtain
Dom δT = (KT∗ )−1 (Dom δTW )
and δT (u) = δTW (KT∗ u) proving the claim.
Now we are ready show that the definition of the multiple Wiener integral
IT,p in the previous section is correct in the sense that it coincides with the
iterated Skorohod integral:
Proposition 4.3. Let h ∈ HT,p . Then h is iteratively p times Skorohod
integrable and
Z T Z T
h(t1 , . . . , tp ) δXt1 · · · δXtp = IT,p (h)
···
0
0
Proof. Again the idea is to use the transfer principle together with induction.
Note first that the statement is true for p = 1 by definition and assume next
that the statement is valid for k = 1, . . . , p. Furthermore, without loss of
generality we can assume that the function h(t1 , . . . , tp ) is of form
h(t1 , . . . , tp ) = f (t1 , . . . , tp−1 )g(tp )
and the general case follows by approximating. Hence, by induction assumption, we have
Z T
Z T Z T
IT,p (f )g(v) δXv .
h(t1 , . . . , tp , tv ) δXt1 · · · δXtp δXv =
···
0
0
0
Put now F = IT,p (f ) and u(t) = g(t). Hence by [14, Proposition 1.3.3] and
by applying the transfer principle we obtain that F u belongs to Dom δT and
δT (F u) = δT (u)F − hDt F, u(t)iHT
W
∗
= ITW (KT∗ g)IT,p
(KT,p
f ) − phIT,p−1 (f (·, t)), g(t)iHT
W
∗
˜ 1 KT∗ g)
= IT (g)IT,p (f ) − pIT,p−1
(KT,p−1
f⊗
˜ KT ,1 g).
= IT (g)IT,p (f ) − pIT,p−1 (f ⊗
Hence the result is valid also for p + 1 by Proposition 4.2 with q = 1.
12
SOTTINEN AND VIITASAARI
We end this section by providing an extension of Itˆo formulas provided by
Al`os, Mazet and Nualart [2]. They considered Gaussian Volterra processes,
i.e., they assumed the representation
Z t
Xt =
K(t, s) dWs ,
0
where the Kernel K satisfied certain technical assumptions. In [2] it was
proved that in the case of Volterra processes one has
Z t
Z
1 t ′′
′
f (Xs ) δXs +
(4.3)
f (Xt ) = f (0) +
f (Xs ) dR(s, s)
2 0
0
if f satisfies the growth condition
2
(4.4)
max |f (x)|, |f ′ (x)|, |f ′′ (x)| ≤ ceλ|x|
−1
for some c > 0 and λ < 14 sup0≤s≤T EXs2
. In the following we will
consider different approach which ables us to:
(i) prove that such formula holds for any polynomial without any additional conditions,
(ii) give more instructive proof of such result,
(iii) extend the result from Volterra context to more general Gaussian
processes,
(iv) drop some technical assumptions posed in [2].
For simplicity, we assume that the variance of X is of bounded variation
to guarantee the existence of the integral
Z T
f ′′ (Xt ) dR(t, t).
(4.5)
0
If the variance is not of bounded variation, then the integral (4.5) may be
understood by integration by parts if f ′′ is smooth enough, or in the general
case, via the inner product h·, ·iHT . In Theorem 4.3 we also have to assume
that the variance of X is bounded.
The result for polynomials is straightforward.
Proposition 4.4 (Itˆo formula for polynomials). Let X be a separable centered Gaussian process with covariance R and assume that p is a polynomial.
Then for each t ∈ [0, T ] we have
Z
Z t
1 t ′′
′
p (Xs ) dR(s, s).
p (Xs ) δXs +
(4.6)
p(Xt ) = p(X0 ) +
2 0
0
Proof. By definition and applying transfer principle, we have to prove that
p′ (X· )1t belongs to domain of δT and that
Z t
(4.7)
Ds GKT∗ [p′ (X· )1t ] ds
E
0
Z
1 t ′′
= E [Gp(Xt )] − E [Gp(X0 )] −
E Gp (Xt ) dR(t, t)
2 0
FREDHOLM REPRESENTATION
13
for every random variable G from a total subset of L2 (Ω). In other words,
it is sufficient to show that (4.9) is valid for random variables of form G =
InW (h⊗n ), where h is a step function.
Note first that it is sufficient to prove the claim only for Hermite polynomials Hk , k = 1, . . .. Indeed, it is well-known that any polynomial can be
expressed as a linear combination of Hermite polynomials and consequently,
the result for arbitrary polynomial p follows by linearity.
We proceed by induction. First it is clear that first three polynomials H0 ,
H1 and H2 satisfies (4.7). Assume next that the result is valid for Hermite
polynomials Hk , k = 0, 1, . . . n. Then, recall well-known recursion formulas
Hn+1 (x) = xHn (x) − nHn−1 (x),
Hn′ (x) = nHn−1 (x).
The induction step follows with straightforward calculations by using the
recursion formulas above and [14, Proposition 1.3.3]. We leave the details
to the reader.
We will now illustrate how the result can be generalized for functions satisfying (4.4) by using Proposition 4.4. First note that the growth condition
(4.4) is indeed natural since it guarantees that the left side of (4.3) is square
integrable. Consequently, since operator δT is a mapping from L2 (Ω; HT )
into L2 (Ω), functions satisfying (4.4) are largest class of functions for which
(4.3) can hold. However, it is not clear in general whether f ′ (X· )1t belongs
to Dom δT . Indeed, for example in [2] the authors posed additional conditions on the Volterra kernel K to guarantee this. As our main result we
show that surprisingly, Ekf ′ (X· )1t k2HT < ∞ implies that (4.3) holds. In
other words, the Itˆo formula (4.3) is not only natural but it is also the only
possibility.
Theorem 4.3 (Itˆo formula for Skorohod integrals). Let X be a separable
centered Gaussian process with covariance R and assume that f ∈ C 2 satisfies growth condition (4.4) and that the variance of X is bounded and of
bounded variation. If
Ekf ′ (X· )1t k2HT < ∞
(4.8)
for any t ∈ [0, T ], then
f (Xt ) = f (X0 ) +
Z
t
f ′ (Xs ) δXs +
0
1
2
Z
t
f ′′ (Xs ) dR(s, s).
0
Proof. In this proof we assume, for notational simplicity and with no loss of
generality, that sup0≤s≤T R(s, s) = 1.
First it is clear that (4.8) implies that f ′ (X· )1t belongs to domain of δT .
Hence we only have to prove that
EhDT G, f ′ (X· )1t iHT
1
= E[Gf (Xt )] − E[Gf (X0 )] −
2
Z
t
0
E[Gf ′′ (Xs )] dR(s, s).
14
SOTTINEN AND VIITASAARI
for every random variable G = InW (h⊗n ).
Now, it is well-known that Hermite polynomials, when properly scaled,
form an orthogonal system in L2 (R) when equipped with the Gaussian measure. Now each f satisfying the growth condition (4.4) have a series representation
∞
X
f (x) =
αk Hk (x).
k=0
Indeed, the growth condition (4.4) implies that
Z
x2
−
|f ′ (x)|2 e 2 sup0≤s≤T R(s,s) dx < ∞.
R
Furthermore, we have
f (Xs ) =
∞
X
αk Hk (Xs )
k=0
where the series converge almost surely and in L2 (Ω), and similar conclusion
is valid for derivatives f ′ (Xs ) and f ′′ (Xs ).
Then, by applying (4.8) we obtain that for any ǫ > 0 there exists N = Nǫ
such that we have
EhDT G, fn′ (X· )1t iHT < ǫ,
where
fn′ (Xs )
=
∞
X
n≥N
αk Hk′ (Xs ).
k=n
Consequently, for random variables of form G = InW (h⊗n ) we obtain, by
choosing N large enough and applying Proposition 4.4, that
Z
1 t
E(Gf ′′ (Xt )) dR(t, t)
E[Gf (Xt )] − E[Gf (X0 )] −
2 0
− EhDT G, f ′ (X· )1t iHT
= EhDT G, fn′ (X· )1t iHT
< ǫ.
Now the left side does not depend on n which concludes the proof.
By Theorem 4.3 above it is straightforward to check in which cases the Itˆo
formula (4.3) is valid. For example, the following is a simple generalization
of [2, Theorem 1].
Corollary 4.1. Let X be a separable centered continuous Gaussian process
with covariance R that is bounded and such that the Fredholm kernel KT is
of bounded variation and
2
Z T Z T
kXt − Xs kL2 (Ω) |KT |(dt, s) ds < ∞.
0
0
FREDHOLM REPRESENTATION
15
Then for any t ∈ [0, T ] we have
Z t
Z
1 t ′′
f ′ (Xs ) δXs +
f (Xt ) = f (X0 ) +
f (Xs ) dR(s, s).
2 0
0
Proof. Note that assumption is a Fredholm version of condition (K2) in [2]
which implies condition (4.8). Hence the result follows by Theorem 4.3. 5. Applications
5.1. Equivalence in Law. The transfer principle has already been used
in connection with the equivalence of law of Gaussian processes in e.g. [17]
in the context of fractional Brownian motions and in [7] in the context of
Gaussian Volterra processes satisfying certain non-degeneracy conditions.
The following proposition uses the Fredholm representation (3.1) to give
a sufficient condition for the equivalence of general Gaussian processes in
terms of their Fredholm kernels.
˜ be two Gaussian process with Fredholm kerProposition 5.1. Let X and X
˜
nels KT and KT , respectively. If there exists a Volterra kernel ℓ ∈ L2 ([0, T ]2 )
such that
Z T
˜
KT (t, u)ℓ(u, s) du,
(5.1)
KT (t, s) = KT (t, s) −
s
˜ are equivalent in law.
then X and X
Proof. Recall that by the Hitsuda representation theorem [9, Theorem 6.3]
˜ is equivalent in law to a Brownian motion
a centered Gaussian process W
on [0, T ] if and only if there exists a kernel ℓ ∈ L2 ([0, T ]2 ) and a Brownian
˜ admits the representation
motion W such that W
Z tZ s
˜ t = Wt −
(5.2)
W
ℓ(s, u) dWu ds.
0
0
Let X have the Fredholm representations
Z T
KT (t, s) dWs .
(5.3)
Xt =
0
˜ is equivalent to X if it admits, in law, the representation
Then X
Z T
d
˜ s,
˜
KT (t, s) dW
(5.4)
Xt =
0
˜ is connected to W of (5.3) by (5.2).
where W
In order to show (5.4), let
˜t =
X
Z
0
T
˜ T (t, s) dW ′
K
t
16
SOTTINEN AND VIITASAARI
˜ Here W ′ is some Brownian motion.
be the Fredholm representation of X.
Then, by using the connection (5.1) and the Fubini theorem, we obtain
Z T
˜ T (t, s) dWs′
˜t =
K
X
0
Z
T
T
=
Z
T
=
Z
Z
T
T
=
Z
T
=
Z
Z
T
d
=
0
0
˜ T (t, s) dWs
K
Z
KT (t, s) −
T
s
KT (t, s) dWs −
0
=
KT (t, s) dWs −
0
KT (t, s) dWs −
0
Z
Z
Z
KT (t, s) dWs −
0
=
KT (t, u)ℓ(u, s) du dWs
0
TZ T
0
TZ s
KT (t, u)ℓ(u, s) du dWs
s
KT (t, s)ℓ(s, u) dWu ds
Z s
ℓ(s, u) dWu ds
KT (t, s)
0
T
0
0
Z
s
ℓ(s, u) dWu ds
0
˜ s.
KT (t, s) dW
0
Thus, we have shown the representation (5.4), and consequently the equiv˜ and X.
alence of X
5.2. Series Expansions. The Mercer square root (3.5) can be used to build
the Karhunen–Lo`eve expansion for the Gaussian process X. But the Mercer
form (3.5) is seldom known. However, if one can find some kernel KT such
that the representation (3.1) holds, then one can construct a series expansion
for X by using the transfer principle of Theorem 4.1 as follows:
Proposition 5.2 (Series expansion). Let X be a separable Gaussian process with representation (3.1). Let (φTj )∞
j=1 be any orthonormal basis on
2
L ([0, T ]). Then X admits the series expansion
∞ Z T
X
φTj (s)KT (t, s) ds · ξj ,
(5.5)
Xt =
j=1
0
where the (ξj )∞
j=1 is a sequence of independent standard normal random variables. The series (5.5) converges in L2 (Ω); and also almost surely uniformly
if and only if X is continuous.
The proof below uses reproducing kernel Hilbert space technique. For
more details on this we refer to [8] where the series expansion is constructed
for fractional Brownian motion by using the transfer principle.
Proof. The Fredholm representation (3.1) implies immediately that the reproducing kernel Hilbert space of X is the image KT L2 ([0, T ]) and KT is
actually an isometry from L2 ([0, T ]) to the reproducing kernel Hilbert space
FREDHOLM REPRESENTATION
17
of X. The L2 -expansion (5.5) follows from this due to [1, Theorem 3.7]
and the equivalence of almost sure convergence of (5.5) and continuity of X
follows [1, Theorem 3.8].
References
[1] R.J. Adler. An Introduction to Continuity, Extrema, and Related Topics for General
Gaussian Processes, volume 12 of Institute of Mathematical Statistics Lecture Notes–
Monograph series. Institute of Mathematical Statistics, Hayward, CA, 1990.
[2] E. Al`
os, O. Mazet, and D. Nualart. Stochastic calculus with respect to Gaussian
processes. The Annals of Probability, 29(2):766–801, 2001.
[3] E. Azmoodeh, T. Sottinen, L. Viitasaari, and A. Yazigi. Necessary and sufficient
conditions for H¨
older continuity of Gaussian processes. Statist. Probab. Lett., 94:230–
235, 2014.
[4] C. Bender and R. Elliott. On the Clark–Ocone theorem for fractional Brownian motions with Hurst parameter bigger than a half. Stoch. Stoch. Rep., 75(6):391–405,
2003.
[5] A. Dasgupta and G. Kallianpur. Chaos decomposition of multiple fractional integrals
and applications. Probability Theory and Related Fields, 115(4):527–548, 1999.
[6] A. Dasgupta and G. Kallianpur. Multiple fractional integrals. Probability Theory and
Related Fields, 115(4):505–525, 1999.
[7] Baudoin F. and Nualart D. Equivalence of Volterra processes. Stochastic processes
and their applications, 107:327–350, 2003.
[8] H. Gilsing and T. Sottinen. Power series expansions for fractional Brownian motions.
Theory of Stochastic Processes, 9(25):38–49, 2003.
[9] T. Hida and M. Hitsuda. Gaussian Processes, volume 120. Translations of Mathematical Monographs, AMS, Providence, 1993.
[10] S. Huang and S. Cambanis. Stochastic and multiple Wiener integrals for Gaussian
processes. The Annals of Probability, 6(4):585–614, 1978.
[11] K. Ito. Multiple Wiener integral. J. Math. Soc. Japan, 3(1):157–169, 1951.
[12] Yu. Mishura. Stochastic calculus for fractional Brownian motion and related processes,
volume 1929 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 2008.
[13] I. Nourdin and G. Peccati. Stein’s method and exact Berry-Esseen asymptotics for
functionals of Gaussian fields. The Annals of Probability, 37(6):2231–2261, 2009.
[14] D. Nualart. The Malliavin Calculus and Related Topics. Probability and Its Applications. Springer, 2006.
[15] V. Perez-Abreu and C. Tudor. Multiple stochastic fractional integrals: a transfer principle for multiple stochastic fractional integrals. Bol. Soc. Mat. Mexicana, 8(3):187–
203, 2002.
[16] V. Pipiras and M. Taqqu. Are classes of deterministic integrands for fractional Brownian motion on an interval complete? Bernoulli, 7(6):873–879, 2001.
[17] T. Sottinen. On Gaussian processes equivalent in law to fractional Brownian motion.
J. Theoret. Probab., 17(2):309–325, 2004.
Tommi Sottinen, Department of Mathematics and Statistics, University of
Vaasa, P.O. Box 700, FIN-65101 Vaasa, FINLAND
E-mail address: [email protected]
Lauri Viitasaari, 1) Department of Mathematics and System Analysis, Aalto
University School of Science, Helsinki, P.O. Box 11100, FIN-00076 Aalto, FINLAND
18
SOTTINEN AND VIITASAARI
¨ cken, Post2) Department of Mathematics, Saarland University, Saarbru
¨ cken, GERMANY
fach 151150, D-66041 Saarbru
E-mail address: [email protected]